Page 40 - IJAMD-1-1

International Journal of AI
for Material and Design ML for quality improvement in L-PBF

that when predicting width, depth, and area within the in scan velocity. This finding highlights laser power as
substrate, ML models demonstrate relatively enhanced the most crucial factor in controlling porosity in these
predictability, achieving the highest R value of more processes.
2
than 0.9. Notably, ML models such as NN, RF, and SVM
generally exhibit good predictive performance. However, 3.1.3. Hardness
the predictability of ML models for height and area based In the context of L-PBF printing, the hardness of the
on height is relatively lower. This discrepancy is attributed manufactured component is a crucial characteristic.
to the complexity involved in predicting the dynamic Materials with elevated hardness exhibit superior wear
powder motions, whereby the models lack the capability to resistance and durability. Hardness, being a measure
react instantaneously to environmental changes. of the material’s mechanical properties, plays a pivotal
In addition, other ML methods, such as the multideity role in determining the quality and performance of the
46
Gaussian process and the neighboring-effect modeling manufactured components. One significant factor
method, have demonstrated good predictive accuracy for affecting the hardness of the manufactured parts is the
determining the geometry of the melt pool. This superior choice of powder material. For instance, the choice of steel
capacity indicates a promising direction for future research types or alloys can directly influence the resultant hardness.
in the application of ML techniques to the field of materials Within the L-PBF printing process, the hardness of the
science, particularly in predicting and understanding the manufactured parts experiences fluctuations in response
dynamics of melt pool formation. 39,40 to variations in input parameters. Therefore, maintaining
stringent control and monitoring of these parameters
3.1.2. Porosity during the manufacturing process becomes essential to

The porosity of a fabricated product, which denotes achieve high-quality printed components.
the extent of voids within it, is one of the critical factors Maitra et al. employed three predictive methodologies,
affecting the quality of L-PBF products. 41 Increased namely Gaussian process regression (GPR), NN, and
porosity levels result in a reduction in the density of parametric multiple linear regression (MLR), for the
the fabricated part, subsequently leading to diminished estimation of hardness in Ti-6Al-4V alloy. In their
47
mechanical performance and increased susceptibility to investigation, the GPR and NN models, following rigorous
brittleness and fracture. Concerning the surface quality optimization and validation through supervised learning
42
of the product, an increased porosity can result in a techniques, exhibited coefficients of determination (R )
2
rough surface or one marred with defects, presenting a of 83% and 90%, respectively, during the training phase.
challenge for components that require meticulous surface These results highlight the critical influence of scanning
refinement. In the face of these problems, the application speed and volumetric energy density on the hardness
43
of ML to reduce porosity emerges as an effective approach. of the fabricated material. In a separate investigation,
Once a trained model establishes the relationship between Ravichander et al. utilized an ANN approach to predict the
L-PBF process parameters and porosity, it opens up the surface hardness of materials. The model incorporated
48
possibility for subsequent porosity reduction. laser power, hatch spacing, and scanning speed as input
variables. The study revealed a significant and inverse
Tapia et al. have developed a Gaussian process-driven
predictive model for the acquisition and forecasting of correlation between hatch spacing and the surface hardness
porosity in metallic components. Their research has of the samples. Therefore, the aforementioned research
44
introduced a Gaussian process-centered predictive model elucidates the impact of different input parameters on the
that characterizes the porosity of the manufactured hardness of the manufactured components.
component as a mathematical function of laser power and In the study conducted by Zhang et al., the author opted
scanning speed. The case study effectively achieved the for RF, XGBoost, and LightGBM as predictive models for
aim of identifying parameter configurations that yield a hardness prediction in samples. Among these models,
49
minimal porosity level of 0.325%, specifically at laser power RF exhibited superior performance in the prediction
(P) = 50 W and scanning speed (v) = 275 mm/s. Imani et al. results, outperforming both XGBoost and LightGBM.
investigated the effect of three process parameters – laser However, when considering training and prediction time,
power, hatch spacing, and scan velocity – on porosity. The along with the capability to sustain excellent performance
45
study revealed that a 50% reduction in laser power, from for subsequent industrial applications while ensuring
340 W to 170 W, significantly increases porosity, leading to accuracy, XGBoost emerged as the overall best performer.
almost three times more pores than an equivalent increase This observation demonstrates that accuracy alone does
in hatch spacing and ten times more than an increase not suffice as the sole criterion for measuring model

Volume 1 Issue 1 (2024) 34 https://doi.org/10.36922/ijamd.2301